• More than US$600 billion in economic damages during the last 20 years
• 721,000 fatalities
• Very hard to predict

• Need for rapid response
• Need for fine-grained spatial resolution
• Not enough data

✅ Parametric insurance. Designed to pay a pre-determined amount for specific pre-agreed scenarios

• Design a catastrophe risk model based on Monte-Carlo simulation

❌ Still no fine-grained spatial resolution

More than 100,000 Monte-Carlo simulations are not enough. Large areas uncovered and very few ‘big’ earthquakes.

• Data doesn’t have to be regularly spaced
• Performs spatial interpolation including covariates (Magnitude, Depth)
• No knowledge of the functional form of the modeled relationshiop is needed

In R, TPS can be adjusted with the Tps function in the fields package. The level of smoothing for the spline must be optimized by cross-validation.

Components of the model:

• Spatial coordinates
• Covariates (Magnitude and Depth)
• Polynomial function for the drift or spatial trend

• Predictions cover all the required area since we are fitting a surface.
• Prediction error is acceptable and calibration is good.

• Thin plate splines are a good alternative for complementing the results of catastrophe risk models based on Monte-Carlo simulations

• They can deal with irregularly spaced spatial data and produce smooth estimates over all the studied area.

• In our model for Chile, prediction error and calibration were satisfactory.

Future work: Implement a methodology to estimate uncertainty in the predictions –> bootstrap